Leukemia diagnosis has never depended on a single test. Morphology remains fundamental, but today it sits alongside flow cytometry, cytogenetics, and increasingly detailed molecular profiling. Together, these tools have made leukemia classification more precise, while also making the diagnostic process more complex.
Artificial intelligence (AI) and digital pathology are beginning to enter this process. Their immediate value is not in replacing the microscope, but in turning some of the information traditionally interpreted visually into measurable, reproducible data.
Algorithms can already identify and classify blood and bone marrow cells, quantify morphologic features, and recognize patterns associated with acute leukemia. The more ambitious goal is to connect those findings with other diagnostic information and help clinicians reach an integrated diagnosis more efficiently.
How AI and digital pathology may change leukemia diagnosis, and where their limitations remain, will be discussed at LeukO 2026, OncoDaily’s global congress on leukemias, taking place virtually on September 3-4.
What Changes When Morphology Becomes Digital?
Examining a peripheral blood smear or bone marrow aspirate generally means looking through a microscope, identifying individual cells, estimating their proportions, and interpreting abnormalities based on the expertise of the hematologist or hematopathologist. That expertise remains indispensable, but the process has inherent variability.
Digital pathology changes what can be done with the same specimen. Once a blood or bone marrow slide is converted into a high-resolution digital image, individual cells can be detected, segmented, measured, and analyzed computationally. Morphology becomes something an algorithm can interrogate at scale.
This has particular relevance in hematology, where diagnosis often depends on recognizing patterns across populations of cells rather than finding a single abnormal structure. AI systems can analyze features such as nuclear shape, chromatin texture, cytoplasmic characteristics, cell size, and other visual patterns across hundreds or thousands of cells. Deep-learning approaches can also learn discriminative image features directly from training data.
What Can AI Actually Recognize in Leukemia?
Acute leukemia has become an important testing ground for AI-assisted morphology. Studies have explored automated blast detection, differentiation of normal and abnormal leukocytes, classification of acute leukemias, and recognition of specific morphologic patterns from peripheral blood and bone marrow images.
Recent reviews of AI-assisted acute leukemia diagnosis show that deep-learning models can achieve high performance in controlled datasets, particularly for tasks involving cell detection and classification. Convolutional neural networks and related architectures have been used to distinguish leukemic from normal cells and, in some studies, to differentiate AML from ALL or recognize particular leukemia subtypes.
An algorithm can screen large numbers of cells, flag suspicious populations, perform preliminary differentials, and direct attention toward abnormal areas. In laboratories processing large specimen volumes, this could reduce the amount of routine visual work while increasing consistency.
Some applications may be particularly valuable when time matters. Acute promyelocytic leukemia (APL), for example, requires rapid recognition because early treatment decisions can be lifesaving. Such a system would not replace molecular confirmation of PML::RARA, but it could potentially help flag a specimen for urgent expert review and testing.
High Accuracy Does Not Equal a Leukemia Diagnosis
Impressive performance numbers require careful interpretation. An algorithm that identifies leukemic blasts with 95% accuracy has solved an important image-recognition problem, but it has not necessarily demonstrated that it can diagnose leukemia with comparable accuracy. Cell classification and patient diagnosis operate at different levels.
A peripheral blood or bone marrow specimen contains a heterogeneous population of cells. A clinically useful system has to move from identifying individual cells to understanding their distribution across an entire specimen.
This distinction is increasingly reflected in AI research. Newer approaches are moving beyond isolated cell classification toward cell-to-patient models, in which predictions across many individual cells are aggregated to produce patient-level estimates.
It also exposes one of the limitations of comparing algorithms based only on classification accuracy. A model can perform extremely well on carefully selected cell images while struggling when confronted with an entire real-world specimen containing staining variation, damaged cells, overlapping cells, unusual morphologies, or rare disease presentations.
Leukemia Diagnosis: Every Step From First Symptoms to Treatment and Beyond
From Image Recognition to Integrated Hematopathology
Leukemia generates several forms of data at the same time: digital morphology, flow cytometry, cytogenetics, molecular sequencing, blood counts, and clinical information. Each describes the disease from a different perspective. Multimodal AI offers the possibility of analyzing these data together. Computational models could eventually assist with the same integration at a scale and speed difficult to achieve manually.
AI is also being explored directly in flow cytometry, including automated gating, identification of abnormal cell populations, and classification of hematologic malignancies from high-dimensional immunophenotypic data. This extends computational diagnosis beyond the digital slide and toward a broader model of computational hematopathology.
When 95% Accuracy Is Not Enough
High performance in a retrospective dataset does not guarantee that an AI system will work equally well in routine practice. This is one of the main barriers between promising research and clinical deployment.
Digital pathology models are sensitive to how images are created. Differences in staining, slide preparation, scanner type, image resolution, laboratory workflow, and patient population can all affect performance. A model trained at one center may therefore perform less reliably when used elsewhere. Small datasets and limited representation of rare leukemia subtypes add another problem: the cases in which expert interpretation is most valuable may also be the ones least represented during model training.
For this reason, external validation matters as much as headline accuracy. A clinically useful system should perform consistently across laboratories, instruments, and patient populations, not only on the dataset from which it was developed. Prospective evaluation is even more important because it shows how the model behaves when incorporated into an actual diagnostic workflow.
The Problem of Bias and the Black Box
Another limitation is that many high-performing AI systems are difficult to interpret. A pathologist may receive a classification or probability score without being able to see which features drove the model toward that conclusion.
In hematopathology, this matters because unusual cases often require understanding why findings do not fit the expected pattern. Explainable AI approaches attempt to show which cells or image regions influenced the prediction, but interpretability remains imperfect. There is also the risk of automation bias, in which a clinician gives excessive weight to an algorithmic result because it appears objective.
Could Artificial Intelligence Reduce Diagnostic Inequality?
Leukemia diagnosis is particularly vulnerable to disparities in expertise. Morphologic interpretation can be difficult, and advanced flow cytometry or molecular testing may not be equally available across regions. Digital slides can already support remote consultation, allowing specimens to be reviewed by experts without physically transporting glass slides.
There is, however, an important limitation. Digital pathology itself requires infrastructure: reliable imaging, standardized slide preparation, data storage, computing resources, technical support, and often stable internet connectivity. The centers that might benefit most from automated support may also have the greatest difficulty implementing it.
Will AI Replace the Hematopathologist?
This question is often framed too broadly. Many tasks within hematopathology are highly repetitive and well suited to automation, others depend on clinical judgment, integration, and interpretation of uncertainty.
AI may increasingly take over parts of the workflow such as cell detection, counting, preliminary classification, and screening of large numbers of images. That could reduce time spent on routine visual tasks and allow specialists to concentrate on the cases in which findings are discordant or difficult to interpret.
These are the situations in which diagnosis depends on deciding which result is technically reliable, which abnormality is meaningful, and whether additional testing is needed.
From Retrospective Models to Clinical Tools
Much of the current literature evaluates AI retrospectively. Prospective studies need to determine whether AI actually improves workflow. Does it shorten turnaround time? Does it reduce disagreement between observers? Does it help identify urgent cases earlier? Does it maintain performance when specimens come from multiple laboratories? Does it change diagnostic or therapeutic decisions in a way that benefits patients?
These endpoints are less visually impressive than a very high accuracy score, but they are more important for clinical implementation.
There is also a practical issue of workflow design. An algorithm that performs well but requires extensive manual preparation, specialized scanners, or difficult software integration may offer little benefit to a busy laboratory. Clinical value depends on how easily the system can be incorporated into routine practice.
Explore More at LeukO 2026
Artificial intelligence and digital pathology are already changing how blood and bone marrow images can be analyzed, but the field remains at an intermediate stage. Cell recognition and classification have advanced rapidly, while generalizability, external validation, workflow integration, and multimodal interpretation remain active areas of development.
How these technologies should be introduced into clinical hematopathology, and how their performance should be judged beyond accuracy alone, will be part of the discussion at LeukO 2026.
